| Abstract [eng] |
The increasing intensity in maritime traffic pushes the requirement for a better prevention-oriented incident management system. The methods developed and studied in the dissertation improve Maritime situation awareness (MSA) by improving detection of vessel traffic anomalies. Big Data of maritime vessel traffic data does not have pre-labeled anomalies. Therefore, the dissertation describes developed and studied unsupervised learning methods of vessel traffic anomaly detection. These methods are compared with partially supervised learning methods using vessel traffic data integrated with meteorological data from the Lithuanian port of Klaipeda and the Fehmarnbelt region of Denmark. In order to compare, a new method of partially supervised learning was developed based on the integration of a self-organizing map (SOM) with virtual pheromone and replicated method of SOM integration with the gaussian mixtures method. Two new unsupervised learning techniques have been developed to detect anomalies in ship traffic trajectories based on recurrent neural LSMT networks, modified to learn the prediction region. Ship traffic outside the forecast region is classified as anomalous. LSTM prediction learning and LSTM "wild bootstrapping" methods are developed and studied based on this concept. The study observed that SOM integrated with virtual pheromone works better with smaller datasets. LSTM-based methods detect more anomalous vessel trajectories than SOM-based methods, such as a sharp maneuver or unexpected stop. |
